I am currently implementing DeepSpeech for my language. I have 2 directories: train and test. Train has approximately 15000 wavs and test approximately 3000. The problem that I face is during training I have large losses, and for the test part loss is huge.
Unfortunately the words are not recognized as it should be. This is my sh file
Could you please give me some hints on what I should do? Thanks!
The new DeepSpeech PlayBook has sections on both training and testing that can provide some guidance here.
One thing that stands out immediately is that you have no validation set. This would be the first thing I would change - ensuring that you have a validation set. See this article for more information on the difference between test, validation and training sets.
You are using a custom alphabet.txt file which is great - I'm assuming that this contains all the characters of the Romanian alphabet?
The next thing here would be to reducing learning rate on plateau - this is covered in the PlayBook.
Mult noroc.
